Dynamic Modeling And Optimization For Combustion System Of The Boiler In Power Plant Based On Spark

At present,in China's power industry,thermal power generation occupies an absolute dominant position regardless of installed capacity or power generation.At the same time,pollutants such as NO_x emitted by coal combustion become one of the major sources of atmospheric pollutants.As the only thermal energy source in the three major thermal power systems,the combustion system directly affects the energy efficiency of the entire plant.Therefore,the boiler combustion optimization has important research significance.The application of intelligent algorithms to boiler combustion optimization is one of the important measures for energy conservation and emission reduction.With the DCS system widely used in power station boilers,more and more operating data is saved,which provides the possibility for subsequent data analysis and mining,but at the same time,a large amount of operational data also puts higher requirements on traditional machine learning algorithms.That is,an algorithm is required to be able to process a large amount of data.In addition,in the boiler optimization process,the coal quality of the coal used for power generation during each time segment is different,and with the aging of the equipment,the inherent parameters in the equipment will not be invariable.This puts forward the control algorithm.The higher requirement is that the algorithm needs to be updated regularly over time.However,there is no mature boiler combustion optimization algorithm that can simultaneously deal with massive historical data and can update existing algorithms in real time through online data.This is the main problem to be solved in this paper.This article uses the parallelism of MapReduce to use the Spark big data platform to parallelize the two training stages of the traditional online extreme learning machine(OS-ELM).Through performance optimization,the algorithm can not only handle massive historical data,but also can process massive amounts of online data.The specific work of this article is as follows:(1)For the traditional OS-ELM algorithm,when dealing with massive high-dimension data,there are problems such as slow speed and insufficient memory of single machine.By analyzing the two stages of the OS-ELM algorithm,the process of solving the general inverse matrix process in which the resources are consumed is disassembled.After that,we use the MapReduce idea to parallelize and propose the core algorithm of this paper based on the massive historical data-based Distributed Online Extreme Learning Machine(MHD-OS-ELM).The algorithm is divided into two phases:the offline phase can process massive historical data;and the online phase can handle massive amounts of online data.In order to use the Spark cluster to run the algorithm more efficiently,this paper uses Spark optimization technology to optimize the performance of the parallel algorithm.And tested on the UCI data set to verify the effectiveness of the algorithm.(2)In order to improve the practicality of the algorithm in the industrial field,SparkStreaming technology was introduced in the online phase of MHD-OS-ELM.The real-time data in the boiler operation was collected by Spark Streaming,and then sent to the MHD-OS-ELM batch by batch for online phase calculation.(3)Preprocess the data collected from the power plant,and then establish a NO_x emission and combustion thermal efficiency prediction model based on MHD-OS-ELM.Through experiments on one month's operation data of the power plant,it is verified that both prediction models have good accuracy and have certain real-time tracking capabilities.(4)Based on the two predictive models,the multi-objective problem is transformed into a single-objective problem using the weight coefficient method,and the input parameters of the model are optimized using the particle swarm algorithm RD-PSO based on Spark.In this paper,the method of selecting the optimal solution in the multi-objective predictive control algorithm(MOMPC)is introduced into the weight coefficient method,and the system's recommendation ratio is quantitatively calculated to avoid complex and time-consuming man-made decisions.